Deep Q learning from demonstrations